TITLE: Show that these matrices are invertible for all $p>3$
QUESTION [25 upvotes]: I am working on a paper which will extend a result in my thesis and have boiled one problem down to the following: show that the symmetric matrix $M_p$, whose definition follows, is invertible for all odd primes $p$. Letting $p>3$ be prime and $\ell = \frac{p-1}{2}$, we define
$$M_p = \begin{pmatrix} 2ij - p - 2p\left\lfloor\frac{ij}{p}\right\rfloor\end{pmatrix}_{1\leq i,j\leq \ell}$$
Examples:

For $p=5$ we have $M_5 = \begin{pmatrix} -3 & -1 \\ -1 & 3 \end{pmatrix}$ and $\det(M_5) = -1\cdot 2\cdot 5$.

For $p=7$ we have $M_7 = \begin{pmatrix} -5 & -3 & -1 \\ -3 & 1 & 5 \\ -1 & 5 & -3 \end{pmatrix}$ and $\det(M_7) = 2^2 \cdot 7^2$.

For $p=11$ we have $M_{11} = 
\begin{pmatrix} 
-9 & - 7 & -5 & -3 & -1 \\
-7 & -3 & 1 & 5 & 9 \\
-5 & 1 & 7 & -9 & -3 \\
-3 & 5 & -9 & -1 & 7 \\
-1 & 9 & -3 & 7 & -5
\end{pmatrix}$ and
$\det(M_{11}) = -1\cdot 2^4\cdot 11^4$.


Though this (seemingly) nice formula that we see above fails for primes greater than 19, though the determinant has been checked to be non-zero for primes less than 1100. (My apologies if this question is not as motivated or as well discussed as is desired. If there are any questions or if further clarification is needed just let me know!)

REPLY [7 votes]: Here are some elements to solve the question.
1st step. Extend $M_p$ as a $(2\ell)\times(2\ell)$-matrix $N_p$, with the same definition of entries. By the way the entries may be writen as $a_{j,k}=p(2\left\{\frac{jk}p\right\}-1)$ where $\{x\}=x-\lfloor x\rfloor$ is the fractional part of $x$. Remark that $a_{j,p-k}=-a_{j,k}$ and likewise $a_{p-j,k}=-a_{j,k}$. Remark that $a_{j,k}$ depends only upon the product $jk\mod p$, thus can be written $a_{jk}$ instead.
Therefore
$$N_p=\begin{pmatrix} M_p & -M_pF \\ -FM_p & FM_pF \end{pmatrix}=\binom{I}{-F}M_p\begin{pmatrix} I & -F \end{pmatrix}$$
where $F$ is the anti-unit matrix, $f_{j,k}=\delta_j^{\ell+1-k}$.
2nd step. From the formula above, $N_p$ and $M_p$ have the same rank. Thus $M_p$ is invertible if and only if $N_p$ has rank $\ell$. This rank is unchanged by a permutation of rows. Thus we examine instead the matrix $N_p'$ obtained by the involution of rows given by the inversion $j\mapsto j^{-1}$ $\mod p$. That way, the entries of $N_p'$ are the numbers $a_{j^{-1}k}$.
3rd step. The matrix $N_p'$ is diagonalizable, its eigenpairs being explicit, because this is a group matrix (entries of the form $a_{g^{-1}h}$) for the multiplicative group ${\mathbb F}_p^\times$ (the non-zero elements of ${\mathbb Z}/p$). Recall that because ${\mathbb Z}/p$ is a field, this group is isomorphic to the additive group ${\mathbb Z}/2\ell$ : there exists an element $\theta\in{\mathbb Z}/p$ of order exactly $2\ell$, and an isomorphism is $\psi(k)=\theta^k$ ($k\mod2\ell$, $\theta^k\mod p$).
The eigenvectors $v^\omega$ have coordinates
$$v^\omega_k=\omega^{\psi^{-1}(k)},\qquad 1\le k\le2\ell,$$
where $\omega$ is any complex solution of $\omega^{2\ell}=1$. The corresponding eigenvalue is
$$\lambda_\omega=\sum_{r=0}^{2\ell-1}a_{\psi(r)}\omega^r=:Q_p(\omega).$$
4th step. There remains to see that among these $2\ell$ eigenvalues, $\ell$ of them only vanish, so that $N_p$ has rank $2\ell-\ell=\ell$. I leave this as an open question (though I am rather optimistic). At least, the fact that $\ell$ of them vanish is clear, because we have $\theta^\ell=-1$, hence $\psi(k+\ell)=-\psi(k)$. There follows that $a_{\psi(k+\ell)}=-a_{\psi(k)}$, so that $Q_p$ factorizes:
$$Q_p(X)=(X^\ell-1)R_p(X),\qquad R_p(X)=\sum_{r=0}^{\ell-1}a_{\psi(r)}X^r.$$
Thus $\lambda_\omega=0$ for every root of $\omega^\ell=1$. Thus there remains only to prove that if instead $\omega^\ell=-1$ (the other roots of unity), one has $R_p(\omega)\ne0$. I have done a few calculations for small prime numbers $p$, which make me optimistic.
Edit. One can conclude at least when $\ell$ is a power of $2$ (that is when $p$ is a Fermat prime). Because then $X^\ell+1$ is irreducible over $\mathbb Q$, while $R_p\in{\mathbb Z}[X]$ has degree $\ell-1$, hence they cannot share a root.
In the general case, the question reduces to whether some cyclotomic polynomial $\Phi_m$ with $m|\ell$ can be such that $\Phi_m(-X)|R_p(X)$. Notice that $\Phi_m$ is reciprocal, while $R_p$ is not, thus this divisibility will imply some other ones, ...